cudaFlow Algorithms » Parallel Sort

Sort a Range of Items

tf::cudaFlow::sort performs an in-place parallel sort over a range of elements specified by [first, last) using the given comparator. The following code sorts one million random integers in an increasing order on a GPU.

const size_t N = 1000000;
int* vec = tf::cuda_malloc_shared<int>(N);  // vector

// initializes the data
for(size_t i=0; i<N; i++) 
  vec[i] = rand();
} 

// create a cudaFlow of one task to perform parallel sort
tf::cudaFlow cf;
tf::cudaTask task = cf.sort(vec, vec + N);
cf.offload();

assert(std::is_sorted(vec, vec+N));

You can specify a comparator to tf::cudaFlow::sort to alter the sorting order. For example, the following code sorts one million random integers in an decreasing order on a GPU.

const size_t N = 1000000;
int* vec = tf::cuda_malloc_shared<int>(N);  // vector

// initializes the data
for(size_t i=0; i<N; i++) 
  vec[i] = rand();
} 

// create a cudaFlow of one task to perform parallel sort
tf::cudaFlow cf;
tf::cudaTask task = cf.sort(
  vec, vec + N, [] __device__ (int a, int b) { return a > b; }
);
cf.offload();

assert(std::is_sorted(vec, vec+N, [](int a, int b){ return a > b; }));

Miscellaneous Items

Parallel sort algorithms are also available in tf::cudaFlowCapturer with the same API.